Paper Type
Short
Paper Number
PACIS2026-1488
Description
Legal judgment prediction (LJP) has emerged as an important application of AI-enabled decision support in judicial systems. However, existing research primarily focuses on criminal cases and often relies on structured fact summaries that differ from the raw materials used in real adjudication. This study proposes Lawgic, a retrieval-augmented multi-agent framework designed for civil legal judgment prediction. The system decomposes adjudication into structured reasoning stages and integrates two complementary approaches: precedent-based case reasoning and rule-based reasoning structured under the IRAC framework. Lawgic operates directly on trial transcripts and dynamically identifies relevant legal provisions, generating interpretable intermediate reasoning traces aligned with judicial workflows. The framework also incorporates a non-parametric learning mechanism that accumulates reusable reasoning artifacts across cases. Experiments on labor dispute arbitration cases show that Lawgic outperforms several LLM-based baselines in judgment prediction and law article identification, demonstrating the potential of multi-agent AI systems for supporting transparent and reliable judicial decision-making.
Recommended Citation
Li, Ruoxin; Chen, Hailiang; and Miao, Chunyu, "Lawgic: A Retrieval-Augmented Multi-Agent System for Civil Legal Judgment Prediction" (2026). PACIS 2026 Proceedings. 6.
https://aisel.aisnet.org/pacis2026/ai_ml/ai_ml/6
Lawgic: A Retrieval-Augmented Multi-Agent System for Civil Legal Judgment Prediction
Legal judgment prediction (LJP) has emerged as an important application of AI-enabled decision support in judicial systems. However, existing research primarily focuses on criminal cases and often relies on structured fact summaries that differ from the raw materials used in real adjudication. This study proposes Lawgic, a retrieval-augmented multi-agent framework designed for civil legal judgment prediction. The system decomposes adjudication into structured reasoning stages and integrates two complementary approaches: precedent-based case reasoning and rule-based reasoning structured under the IRAC framework. Lawgic operates directly on trial transcripts and dynamically identifies relevant legal provisions, generating interpretable intermediate reasoning traces aligned with judicial workflows. The framework also incorporates a non-parametric learning mechanism that accumulates reusable reasoning artifacts across cases. Experiments on labor dispute arbitration cases show that Lawgic outperforms several LLM-based baselines in judgment prediction and law article identification, demonstrating the potential of multi-agent AI systems for supporting transparent and reliable judicial decision-making.
Comments
01-AIML